Abstract: Document-level event extraction is an event extraction task for long texts with multiple sentences.Existing document-level event extraction studies generally divide event extraction into three sub-tasks:candidate entity extraction,event detection and argument recognition,and usually train them with joint learning.However,most of the existing document-level event extraction methods extract candidate entities sentence-by-sentence without considering the cross-sentence contextual information,which obviously reduces the accuracy of entity extraction and argument recognition.Furthermore,it affects the final event extraction results.Based on this,this paper proposes a document-level event extraction method based on multi-granularity entity heteroge- neous graphs.This method uses two independent encoders,Transformer and RoBerta,for sentence-level and paragraph-level entity extraction respectively.Meanwhile,this paper proposes a multi-granularity entity fusion strategy to select entities that are more likely to be event arguments from the set of sentence entities and paragraph entities,and further constructs a heterogeneous graph incorporating multi-granularity entities.Finally,we use graph convolutional network to obtain document-aware entity and sentence representations for multi-label classification of event types and event arguments to achieve event detection and arguments recognition.Experiments on ChFinAnn and Duee-fin datasets show that the proposed method improves about 1.3% and 3.9% by F1 compared with previous methods,which proves its effectiveness.

Key words: Document-level event extraction, Event extraction, Heterogeneous graph, Entity extraction, Multi-granularity